Evaluation Rules

sativum.ai helps you read food labels with more context: each product gets a Clean Score, ingredients get a Health Score, and the overall product receives one of three verdicts — Recommended, Situational or Not recommended. The rules below are deterministic and apply identically to every product, regardless of brand.

This is an informational guide, not medical, regulatory or nutritional advice. See our Privacy Policy and Terms of Service for the full disclaimer.

The Clean Score measures label transparency and processing load — not calories, not macros. A high score means short, recognisable ingredient lists with few additives; a low score means the opposite.

The score starts from an AI label-reading estimate (0–100) and is then passed through a deterministic cap-and-stack pipeline. Every additive detected is looked up in an evidence-based registry that assigns a function class (sweetener, colour, preservative, flavour enhancer, emulsifier…) and a risk class (high / moderate / low / neutral) with a citation (EFSA re-evaluation, IARC monograph, Southampton study, JECFA opinion). We flag by function and evidence — not by raw presence: neutral natural colours (riboflavin, carotenes, curcumin, annatto, beetroot red) and low-risk technical aids (citric/ascorbic acid, lecithin, gums, pectin) never trigger a cap.

1. AI score — initial 0–100 read from the ingredient list.
2. Caps fire — each red-flag trigger imposes a maximum allowed score (see table below).
3. Top-3 amplification — if a trigger ingredient sits in pack positions 1–3 (i.e. it is one of the main ingredients by mass), its cap tightens by an extra −10 points.
4. Stack penalty — every additional cap beyond the first subtracts another −5 points (max −20).
5. Final score = min(AI score, tightest cap) + stack penalty, clamped to 0–100.

Every community product exposes a collapsible “Why this score?” panel that shows the AI score, every triggered cap (with a top-3 badge when amplified), the stack penalty, and the final number — so the calculation is fully auditable.

Caps in force

Footnote — what does “cap” mean? A cap is a ceiling, not a subtraction. When a trigger fires, the product's score cannot exceed that number, regardless of what the AI initially gave it. Example: AI scores a product 78, but a flavour enhancer is detected → the score is pulled down to 55 (the cap). If multiple triggers fire, the lowest cap wins, and each additional trigger removes another −5 points via the stack penalty (max −20). The top-3 badge means the offending ingredient sits in pack positions 1–3 (a main component by mass), which tightens the cap by another −10.

Each ingredient has a Health Score on a −10 to +10 scale, aggregated from peer-reviewed beneficial and harmful health effects weighted by intensity. Positive = net beneficial effects dominate; negative = net harmful effects dominate.

• Beneficial effects (antioxidant, anti-inflammatory, cardio-protective…) add positive points.
• Harmful effects (pro-inflammatory, allergenic, toxicity at dose…) subtract points.
• Each effect is weighted by its documented intensity in our database.

Product-level Health Score (community products)

For a finished product (a blend of many ingredients), single-ingredient scores are combined with a position-weighted formula — ingredients listed first count for much more than ingredients listed last, because EU labelling rules require descending order by mass.

1. Harmonic position weight: ingredient at index i gets weight wi = 1 / (i + 1). So position 1 = 1.00, position 2 = 0.50, position 3 = 0.33, position 10 = 0.09.
2. Bad-actor amplification: harmful ingredients (negative score) in positions 1–3 are multiplied by 1.5× — a flavour enhancer as the second ingredient hurts the product score much more than the same enhancer at the end.
3. Stacked product penalties add on top of the weighted average: flavour enhancers, palm oil, ≥3 additives, high salt and high added sugar each subtract additional fixed points from the product Health Score.
4. Final score = weighted average of ingredient scores + product penalties, clamped to −10 … +10. A hard ceiling of +1 applies whenever a severe lever fires: salt ≥ 10 g/100 g, sugar ≥ 22.5 g/100 g, or salt / free sugar as the first pack ingredient. Such products can never display "Very Good" or "Excellent".

Each product page exposes a collapsible “Why this score?” panel listing every ingredient with its position weight, its individual contribution and every penalty applied, so the result is fully traceable.

The Health Score reflects general evidence at typical culinary doses. It is not tailored to individual conditions, allergies or medications — always consult a qualified professional for personal advice.

The Culinary Score is orthogonal to Clean and Health. It answers a different question: is this a good example of what it claims to be? Rewards sensory density and cohort standing so an aged cheese, a fish sauce or a curry paste is not punished for being intense — those products have to be intense to work.

How it's computed

1. Richness (0–60 pts) — sum of the top intensities across taste, aroma, mouth-stimulus and texture. Top-3 tastes are weighted ×1.5 because headline flavors define character. Normalized against a reference ceiling of 80 raw points.
2. Cohort standing (0–40 pts) — percentile of this product's richness inside its subcategory. When cohort data is missing the component is skipped (the score then maxes out at 60).
3. Final = richness + cohort, clamped 0–100. Label from Exceptional / Great / Solid / Mild / Bland.

A high Culinary Score does not imply clean or healthy. Always read the three scores together — that is the whole point of keeping them independent.

The Authenticity Score answers a specific question: how faithful is this industrial product to the canonical traditional recipe of the dish it evokes? It only appears when the product has been matched to a canonical dish (Ragù alla Bolognese, Hummus, Pesto Genovese, Guacamole, Tzatziki…). Products with no traditional counterpart don't get this KPI — the tile and star weight are dropped, not penalized.

How it's computed

Score = 60 · coreCoverage + 25 · optionalCoverage + 15 − deviationPenalty, clamped 0–100. Higher-tier canonical sources (PDO / PGI legal text, producer-of-origin) win over consensus writeups when the canonical recipe is researched.

1. Core coverage (0–60 pts) — weighted fraction of the dish's core ingredients that are present on the label. Missing a core ingredient (e.g. tahini in hummus, basil in pesto) is the biggest hit.
2. Traditional optional (0–25 pts) — fraction of the traditional but non-mandatory ingredients present. Missing these is fine, having them helps.
3. Baseline (+15 pts) — everyone starts here; the penalties below eat into it.
4. Deviation penalties — capped per category so no single deviation zeroes the score:
   • Additives / E-numbers / flavourings — up to −30.
   • Added sugars where the traditional recipe has none — up to −20.
   • Wrong oil / fat basis (e.g. sunflower oil instead of olive) — up to −10.
   • Dominant swap: non-canonical ingredient in pack positions 1–3 — up to −15.

Authenticity is orthogonal to Clean, Health and Culinary — a heavily adapted product can still be clean, or a traditional one can still be salt-heavy. It is folded into the Sativum star rating as an equal-weight 7th KPI whenever a canonical match exists.

Every scan is checked for organic evidence on the label — logo text, ingredient list, brand and declared claims — and classified into one of three trust tiers. A leaf badge is surfaced on the community grid tile, the product hero, the compare bar and the PDF export. Certified and Likely tiers add a bonus to the Clean Score (applied after all caps and hard-capped at 100).

Detection sources

• EU Bio-Logo + control code matched against the organic_control_bodies table — code and certifier name are stored on the observation and rolled up to the product.
• Certifier keywords — a curated allow-list of trusted labels (see tier 2 above).
• Loose marketing keywords — multilingual (EN/DE/FR/IT/ES/RO/RU) with a small denylist to avoid false positives like "eco-friendly packaging" or "biological process".

Rollup & bonus

1. Detection runs per observation — the highest tier across all scans of a product wins and is written to aggregate_stats.organic.
2. The bonus (+10 / +5 / 0) is added after the Clean Score cap-and-stack pipeline, then clamped to 100 so it can never override a hard cap into an impossible number.
3. Fake-claim / greenwashing suspects keep the +0 bonus and surface an amber pill so the user sees the mismatch immediately.

The colored chips on each product page are deterministic. Given the three scores plus additive count, the same product always produces the same chips. Rules are evaluated in the priority order below and a product shows at most two chips.

Every community product receives one of three verdicts. The rule is deterministic: we compute it first, then the AI-written narrative is required to be consistent with it.

The bullet-point Strengths and Watch-outs shown on every community product page (and in the Compare view) are not free-form AI text. They come from a deterministic rule registry evaluated against the parsed label: if the signal is true, the callout fires — every time, on every product. The AI's own writeup is only kept when it doesn't duplicate a rule, which is why two products with the same label always surface the same positives and the same negatives.

Strengths (max 6)

Watch-outs (max 6)

The strengths and watch-outs for the same axis are mutually exclusive — a product either gets No palm oil or Contains palm oil, never both. Salt tiers are also mutually exclusive: the highest applicable tier wins. Cap reasons from the Clean Score (see above) are merged into the watch-out list so the UI always explains why the score was capped.

These triggers automatically cap or lower a product's verdict. They are not moral judgements — they exist because each one materially changes what the product is.

The Sativum star rating blends up to seven KPIs (Health, Clean, Nutrition, Culinary, Allergens, Daily-use, Authenticity). Rating a bottle of wine, a bottle of olive oil and a jar of tomato sauce with the same equal weights is misleading — some KPIs are structurally not meaningful for certain product types (e.g. per-100 g Nutri-Grade for pure oils always resolves to E).

We therefore first infer an evaluation archetype from the subcategory (with light heuristics like ABV %, "beer" / "cider" mentions, alcoholic-beverage keywords) and apply a per-archetype weight profile. Suppressed KPIs are dropped from the weighted mean and shown as informational ("not weighted") rows in the explainer so nothing is hidden.

Numbers are nominal weights — renormalized to 100 % across the KPIs that are actually present for a given product. "—" means the KPI is suppressed for that archetype and does not contribute to the weighted mean; it still appears in the explainer as informational.

Worked example — table wine (12.5 % ABV). Health, Clean, Culinary, Allergens and Authenticity contribute; Nutri-Grade and Daily-use are suppressed. If the wine scores +4 Health, 70 Clean, 78 Culinary, "Free" Allergens (100), and "Faithful" 82 Authenticity, the weighted index is (60·15 + 70·20 + 78·25 + 100·10 + 82·30) / 100 ≈ 78 → 3.9★. The same wine under equal weighting would drag in a "Grade E" nutrition tile (20/100) and a null Daily-use, misrepresenting the product.

The qualitative panel on each product page — Sensory character, Culinary use, Best for, Skip if — is generated in two steps:

1. The deterministic verdict above is computed first.
2. An AI narrative is generated and required to stay consistent with the verdict — it can never contradict it.

This keeps the prose readable while guaranteeing the badge and the text always tell the same story.

• Community contributions — label scans and product entries submitted by users.
• Ingredient database — sativum.ai's curated chemistry, sensory and health-effect dataset.
• Multi-model AI pipeline — Perplexity, OpenAI, x.AI and Gemini are used in sequence for descriptive copy, with deterministic logic on top for any scoring decision.
• Scores and verdicts recompute as the underlying ingredient data is refined.

• Source — a strict-JSON LLM pass produces each cuisine's description, cultural notes, dietary traditions, six flavor metrics (spice/heat/umami/fermentation/acidity/sweetness), a signature-ingredient list and a typical-ingredient list.
• Ingredient matching — every AI-suggested ingredient is normalized (diacritics stripped, form descriptors like "ground/dried/fresh" removed) and matched against the canonical ingredients table and its ingredient_name_translations.
• Plausibility guard — cuisines that resolve fewer than 5 ingredients are logged to data_quality_flags for human review instead of being silently saved.
• Auto-relink — a trigger on the ingredients table enqueues every new or renamed ingredient into a relink queue; a scheduled job replays the queue and adds missing links to the cuisines whose stored payload references that ingredient. New ingredients propagate to cuisine coverage without manual work.
• Refresh — admins can re-enrich a single cuisine or bulk-backfill the whole taxonomy from Admin → Cuisines. Runs are timestamped in last_enrichment_at.

sativum.ai is not a certification body. The verdicts on this site are not regulatory approvals, not medical advice, and not a substitute for reading the label yourself. Products evolve; recipes change; our database is improving daily.

If you spot a mistake or want to challenge a verdict, please write to hello@sativum.ai with a photo of the current label — we re-evaluate on request.

Last reviewed: August 2026. Maintained by Manole 13 Labs SRL, Curtea de Arges, Romania.